Experimental assessment of AI-based interactome mapping

Abstract

Genotype-phenotype relationships are mediated through intricate networks of physical and functional interactions among macromolecules. Knowledge of the interactome is vital to understand and model genetics and cellular biology. Recent advances in accurately predicting tertiary protein structures using artificial intelligence (AI) approaches such as AlphaFold1 have revived the vision that the proteinprotein interactome might be fully predictable through computational modeling of quaternary structures. Here we present a comprehensive experimental framework to systematically assess the impact of AIdriven interactome predictions for yeast2 and human3. We find that the quality of high-confidence predictions is on par with established experimental approaches. However, in proteome-wide screening, the tested AI approaches underperform in the discovery of strictly novel protein-protein interactions (PPIs) compared to experimental reference interactome maps. In particular, the yeast interactome map describe here identifies >40-fold more novel PPIs than its AI counterpart. Strikingly, AlphaFold provides structural models for a substantial number of experimentally identified PPIs miss by the virtual screens. Our results suggest that, at this stage, the main contribution of AI predictions is to provide quaternary structure models for experimentally identified PPIs.